Dec 15, 2025

ML Times

• Speech and Language Processing (3rd ed. draft)

  • The Feb 3, 2024 release of "Speech and Language Processing" by Dan Jurafsky and James H. Martin introduces updated chapters and slides, with a promise of more content soon.

• Nemotron 3 Nano

  • Nemotron 3 Nano introduces a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture, achieving 31.6B total parameters and a 1M-token context window, enabling efficient, high-throughput agentic models.

• P-computers can solve spin-glass problems faster than quantum systems

  • UCSB's research reveals that p-computers, utilizing probabilistic bits (p-bits), can outperform quantum systems in solving spin-glass problems, showcasing their potential as a practical alternative for complex optimization tasks.

• From sci-fi to reality: Researchers realize quantum teleportation using tech

  • German scientists have successfully teleported information using quantum entanglement between two independent light-emitting devices, demonstrating that quantum teleportation can leverage existing internet technology without the need for new infrastructure.

• Why proteins fold and how GPUs help us fold

  • NVIDIA's GPUs have evolved from gaming to revolutionizing protein design, enabling the creation of novel proteins that can tackle environmental issues and diseases, showcasing the power of AI in biological research.

• How to Fine-Tune an LLM on NVIDIA GPUs With Unsloth

  • Fine-tuning LLMs with Unsloth on NVIDIA RTX GPUs accelerates the customization of AI models, enabling the creation of personalized assistants for various tasks, including study and work, while leveraging the new Nemotron 3 family of open models for enhanced efficiency.

• StructOpt: a first-order optimizer driven by gradient dynamics

  • StructOpt leverages the sensitivity of the gradient to parameter changes, providing a novel approach to optimization that avoids second-order computations while enhancing stability and adaptability in various landscapes.

• Bounding Hallucinations: Information-Theoretic Guarantees for RAG Systems via Merlin-Arthur Protocols

  • RAG models can be enhanced by treating retrieval as verifiable evidence through an interactive proof system, improving the reliability of large language models (LLMs) in generating accurate responses.

• Does Less Hallucination Mean Less Creativity? An Empirical Investigation in LLMs

  • LLMs can reduce hallucinations through techniques like Chain of Verification (CoVe), which enhances divergent creativity, while Decoding by Contrasting Layers (DoLa) suppresses it, revealing a complex relationship between factual accuracy and creative output.

• Towards Trustworthy Multi-Turn LLM Agents via Behavioral Guidance

  • Large Language Models (LLMs) can achieve trustworthy multi-turn interactions through a novel task completion framework that incorporates behavioral guidance, enhancing their reliability in complex environments.

• CIP: A Plug-and-Play Causal Prompting Framework for Mitigating Hallucinations under Long-Context Noise

  • CIP is a lightweight and plug-and-play causal prompting framework designed to mitigate hallucinations in large language models by establishing causal relationships among entities, actions, and events, thus enhancing reasoning quality.

• A Scalable Multi-GPU Framework for Encrypted Large-Model Inference

  • Cerium is a multi-GPU framework that enables encrypted inference for large models, achieving performance comparable to ASICs while maintaining accessibility through GPU technology.

• Adaptive Soft Rolling KV Freeze with Entropy-Guided Recovery: Sublinear Memory Growth for Efficient LLM Inference

  • ASR-KF-EGR introduces a reversible soft-freeze mechanism that optimally manages key-value updates for low-importance tokens, enhancing efficiency in large language model generation without the need for training.